Automatic speech recognition (ASR) technology can be used to map audio utterances to textual representations of those utterances. In some systems, ASR involves dynamically adapting acoustic model parameters based on features of a particular speaker's utterances to potentially improve speech recognition performance for the particular speaker. However, these adaptations may be device-specific, and the particular speaker may provide utterances to multiple devices.